Machine Learning Assessment of Spasmodic Dysphonia Based on Acoustical and Perceptual Parameters

Area of Science:

• Speech-language pathology
• Biomedical engineering
• Machine learning in healthcare

Background:

• Adductor spasmodic dysphonia (ADSD) is a focal dystonia causing laryngeal muscle spasms.
• Accurate severity assessment is crucial for effective treatment of ADSD.
• Current assessment relies on subjective perceptual measures.

Purpose of the Study:

• To apply machine learning for objective severity assessment of ADSD.
• To identify relationships between perceptual and acoustic measures in ADSD.
• To develop a potential diagnostic tool for ADSD severity.

Main Methods:

• Collected perceptual indices (GRB scale) and acoustic parameters from 28 female ADSD patients.
• Utilized Local Interpretable Model-Agnostic Explanations (LIME) to analyze feature relationships.
• Developed and validated a k-nearest neighbors (KNN) classification model using cross-validation.

Main Results:

• Established reliable correlations between GRB scores (G, R, B, Spasmodicity) and acoustic parameters (voiced percentage, F2 median, F1 median).
• The KNN model achieved 89% accuracy in classifying ADSD severity into mild, moderate, and severe categories.
• Identified key acoustic parameters crucial for objective ADSD assessment.

Conclusions:

• Machine learning, particularly KNN, can effectively assess ADSD severity.
• Specific acoustic parameters combined with perceptual indices offer a robust approach for ADSD evaluation.
• The findings support the development of objective tools to aid clinical judgment in ADSD management.